#8 Ethics and the Evolution of NLP
Faculty of Humanities and Social Sciences
University of Lucerne
20 April 2023
.txt, incl. OCRCopy full assignment into ChatGPT
Get nicely structured, more or less unique output
However, especially for task 1.3, the answers are plain wrong
Von BRUNO VANONI, BERN.“How do I parse two uppercased words at the beginning of a line after the word ‘Von’?”
Face impressions as perceived by a model by (Peterson et al. 2022)
💡 What is going on behind the scene?
King = 1, Queen = 2, Man = 3, Woman = 4Vector-representations of words as discrete symbols (Colyer 2016)
NLP is great. I love NLP.
I understand NLP.
NLP, NLP, NLP.
NLP |
I |
is |
term | |
|---|---|---|---|---|
| Doc 1 | 2 | 1 | 1 | … |
| Doc 2 | 1 | 1 | 0 | … |
| Doc 3 | 3 | 0 | 0 | … |
| Doc ID | … | … | … | term frequency |
___ for lunch.”You shall know a word by the company it keeps!
Firth (1957)
King – Man + Woman = QueenFrance / Paris = Switzerland / Bernbank)💥 embeddings are the cornerstone of modern NLP
🤓 There are dozens other models than ChatGPT.
«___ becomes a doctor.»
Gender bias of the commonly used language model BERT (Devlin et al. 2019)
Gender bias of the commonly used language model BERT (Devlin et al. 2019)
Simplified illustration of a Neural Network. Arrows are weights.
🤓 train with gradient descent, a series of small steps taken to minimize an error function
„This sentence contains 37 characters.“
„Dieser Satz enthält 32 Buchstaben.“
Gender bias in Google Translate
Raw data is an oxymoron.
Gitelman (2013)
Don’t ask if artificial intelligence is good or fair, ask how it shifts power.
Kalluri (2020)
Don’t be a fool. Be wise, think twice.
Text generation may be used to communicate difficult decisions strategically